🚨 96.8% of KernelBench taken from torch.compile, a 300M drafter at 3.18x, and HF checkpoint → C++ in two commands

Presented by Mend.io  Securing AI Agents, MCP Servers & LLM Apps — a see–fix–protect framework for agents already in your codebase. Get the guide →

Every release below removes an intermediate layer. The ONNX export between a checkpoint and a TensorRT engine. The public endpoint between two agents that want to call each other’s MCP tools. The hand-written kernel between PyTorch and the GPU. The fork you maintained because the agent loop was hard-coded. The sequential decode between a prompt and its output. None of these were the workload. They were what had to run before the workload. Seven teams cut one out this week, and the pattern is clearer read together than one release at a time.

96.8%

of KernelBench beats the compiler

2.67x

mean DSpark decode gain, H100

1,123

Elo on the cloned-voice board

01 · NVIDIA · Inference

NVIDIA Releases TensorRT Model Connect in Public Preview: Hugging Face Checkpoint to Native C++ Inference in Two Commands

Getting a checkpoint into a C++ service was a chain: PyTorch, then ONNX or TorchScript, then TensorRT, then per-model integration glue. Every link had its own export gaps and its own validation. TRTMC deletes the middle of the chain and keeps the two ends.

Two commands produce a versioned .bundle that native profiles execute in C++ with no PyTorch in the runtime path. The July 29, 2026 GB300 snapshot covers 105 profiles across 76 families, and 102 beat their declared reference by more than 5%. Apache-2.0. Release wheels are Linux aarch64 only today, so x86_64 means the Docker source build. NVIDIA also states the project itself — implementations, tuning, tests, docs — was built using OpenAI Codex agents under human review.

Read on Marktechpost →  · GitHub  · Build docs  · License

02 · Google · Agent networking

Meet SAM (Sovereign Agent Mesh): A Zero-Config, Zero-Trust P2P Network for AI Agents

Letting two agents share a tool normally means putting an internal script or an LLM endpoint on the public internet. SAM removes the endpoint instead of guarding it. This is not Segment Anything, and the repo carries an explicit disclaimer that it is not an officially supported Google product.

Three binaries: sam-control-plane, sam-router, sam-node. The control plane verifies an OIDC JWT, translates the claims into Datalog facts, and seals them into a Biscuit token — so a node authorizes a call offline against its own local rules without calling home. Default-deny is absolute: even the discovery catalog system://sam.catalog needs an explicit grant. Apache-2.0, with Helm charts and a Kubernetes guide, but the public mesh is still a beta testnet, so production means self-hosting the control plane.

Read on Marktechpost →  · GitHub  · Docs  · Kubernetes guide

03 · ByteDance Seed + Tsinghua AIR · GPU kernels

ByteDance Seed and Tsinghua AIR Introduces CUDA Agent: A Large-Scale Agentic RL System for CUDA Kernel Generation

Frontier models already write correct CUDA. They write slow CUDA. Base Seed1.6 passes 74.0% of KernelBench and beats torch.compile on only 27.2% of tasks, at 0.69x geomean — on average slower than the compiler it was asked to beat.

Put that model inside a real profiling sandbox and train it for 150 PPO steps at a 131,072-token context and it reaches 98.8% pass and 96.8% faster-than-compile at 2.11x geomean across the 250-task benchmark. On the Level-3 split it sits roughly 40 points above Claude Opus 4.5 (50.0%) and Gemini 3 Pro (52.0%) on faster rate. The ablation is the honest headline: remove the agent loop and 96.8% falls to 14.1%. Weights are closed, but the 6,000-sample dataset, the SKILL.md spec and the reward recipe are public.

Read on Marktechpost →  · Paper  · Project page  · Dataset

Sponsored · Mend.io

Securing AI Agents, MCP Servers & LLM Apps: A Practical Framework

Agents, MCP integrations, and LLM-powered apps are entering codebases faster than security programs can track them — and their behavior is not defined by code alone. This guide gives AI/ML engineers, platform engineers, and AppSec teams a practical see–fix–protect framework, with the checklists and templates to start today.

Inside: the agentic AI attack surface map across five layers, from prompt injection to poisoned MCP tools. A 12-point misconfiguration checklist for agent and MCP permissions. Triage at AI speed — what to automate and what to keep human. Runtime guardrail architectures and prompt hardening for production. Plus an agentic AI maturity self-assessment aligned to NIST AI RMF, OWASP AIMA, ISO/IEC 42001, and the EU AI Act.

Download the white paper → · See what is inside

04 · Liquid AI · Speculative decoding

Liquid AI Releases LFM2.5-DSpark Draft Models That Deliver Up to 3.18x Faster Decoding Without Changing Model Outputs

Decoding runs one token at a time because each token waits on the one before it. A roughly 300M drafter proposes a block of nine, the target verifies the whole block in a single forward pass, and under greedy decoding the emitted sequence is identical to the target running alone. Same output, less waiting.

On one H100 via SGLang, LFM2.5-2.6B goes from 323 to 864 tok/s, a 2.67x mean. The peak is 3.18x on MATH500 for 8B-A1B. The honest range is 1.04x to 3.18x, because speedup tracks acceptance: 8.27 of 10 tokens accepted on MATH500 against 4.02 on GSM8K. On-device MoE is the weak spot, at 1.18x mean on an M4 Max. Multi-tool function calling gains most, at 57% lower latency on 2.6B. The LFM Open License v1.0 is free commercially only under $10M annual revenue.

Read on Marktechpost →  · Model card  · Write-up  · SGLang PR  · License

05 · DeepSeek · Agent harness

DeepSeek AI Releases DeepSeek Harness in Developer Preview: An MIT-Licensed Agent Harness Where Everything is a Plugin

Most harnesses hard-code the layer between a model and the environment it acts in. The agent loop, the tool registry and the session store are fixed, and you extend at whatever hooks the authors left behind. Harness states the opposite in the first lines of its README: everything is a plugin.

Models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI all sit behind Cordis plugin boundaries and swap in configuration, without touching Harness source. Four modes ship: Standard, Code, Minimal and Creator. Every run writes to an append-only session log that records system prompts, reasoning, tool results, subagent scheduling and every context injection — most frameworks log the tool calls only — and resume, fork, search and replay all operate on that one event stream. MIT-licensed, ships as dsh, and v0.1 is a developer preview.

Read on Marktechpost →  · GitHub  · Docs  · Product page  · Cordis

06 · Cartesia · Voice

Cartesia Ships Sonic-3.6: A Streaming TTS Model That Now Leads Both Artificial Analysis Speech Arenas

Winning a voice leaderboard usually means you shipped better voices. The Controlled Voice board clones every model onto the same eight reference voices, which isolates the synthesis engine from the catalog. Sonic-3.6 leads that board, which is the result worth reading.

1,283 Elo on Provider Voice and 1,123 on Controlled Voice, with Sonic-3.5 second and ElevenLabs Eleven v3 third. Artificial Analysis normalizes it at $49.00 per 1M characters, half of Eleven v3 at $100.00 and well above Speechify Simba 3.2 at $10.00 for a 1,240 Elo. It runs on state space models rather than transformers. Read the deployment line carefully: hosted API in beta, no open weights and no Hugging Face repo, and the docs still list Sonic 3.5 as stable. Sub-90ms time-to-first-audio is vendor-stated model latency, not a measured end-to-end round trip.

Read on Marktechpost →  · Project page  · Speech Arena  · Docs  · Pricing

Sponsored on Marktechpost

Meet UPDF: A Lightweight Adobe Alternative Built for the Agentic Era

A model can summarize a 90-page agreement. It will not hand back a clean, signable file. UPDF puts editing, 38-language OCR, 14-format conversion and an AI assistant in the same window as the document. Read it →

The Video Production Stack Now Fits on One Desk: LTX-2.5 Launches as NVIDIA-Accelerated Open Weights World Model

A 10-second clip in 6.8 seconds on-prem on 2x GB200, against 52 to 398 seconds for the closed alternatives listed. Open weights, native multishot, day-one ComfyUI, free under $10M ARR. Read it →

One more thing…

MiniMax · Music

MiniMax Releases MiniMax-Music3: An Open-Weights Music Model Generating Complete Five-Minute Songs From Lyrics and a Structured Caption

A song used to be assembled. Write it, track it, arrange it, mix it, master it. Every one of those was a separate room and a separate bill. MiniMax-Music3 takes two inputs — lyrics carrying section tags like [Verse] and [Chorus], and a structured caption covering metadata, vocals and arrangement — and returns up to five minutes of 32 kHz, 16-bit stereo in one generation.

The Hybrid-LM splits the modeling problem. An 8B Global LLM predicts the first RVQ codebook frame by frame and holds the long-range structure; a 0.6B Local LLM fills the remaining seven codebooks inside each frame. Then the interesting part: synthesis skips the discrete tokenizer decoder entirely. The final hidden states of both models are fused and condition a 2.4B flow-matching module into a latent space that a 123M Flow-VAE decodes. At inference the tokenizer decoder is never loaded.

Two CUDA GPUs via SGLang-Omni, under 24 GB via the diffusers pipeline, or 8 GB with leaf-level group offloading. Commercial use is permitted with ‘MiniMax-Music3’ displayed prominently in the product UI, and any organization above USD 20M in yearly revenue from those products needs separate prior written authorization.

Read on Marktechpost → · GitHub · ComfyUI · License

Sponsor AI Dev Signals

Put your product in front of the people who build with it.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

See one in action — this issue’s placement for Mend.io →